An Unsupervised Machine Learning Technique for Recommendation Systems

Due to the advancement of the Internet, various kinds of data are easily found online which helps users to find useful information which are of their interest. However, the exponential growth of data has caused it to be complex and huge, so it has become difficult to filter valuable information from it. Recommendation systems can help to overcome this issue and give recommendations to the users which matches the area of their interest. Most of the systems rely on a rating prediction algorithm where the items are taken as recommended for a user if the user's predicted rating is high on those items. This research aims to increase the accuracy and reduce the processing time for recommendation using the prediction algorithm based on the unsupervised machine learning method. The proposed solution consists of Autoencoder to enhance the accuracy of prediction and reduce the processing time. Partially observed interaction matrix is used as input for the neural network model which outputs a complete rating matrix. The proposed solution achieved an improvement by 1.83%, 0.85% and 3.72% in cold start case for MSE, RMSE and MAE evaluation metrics respectively. The proposed solution performs better and will be used in cold start cases for datasets where timestamp value (user creation time) is used.